Aiming at the problem that grid-connected photovoltaic (PV) system has strict requirements for fault diagnosis accuracy and response time, an online intelligent diagnosis method for grid-connected PV system faults based on MAFDT is proposed. This study enhances PV fault diagnosis by combining CBDF-based multi-agile features with a DT in the MAFDT framework and deploying DSP-based online diagnosis to reduce response time and improve reliability. Firstly, the output data of the PV system is subjected to multi-agile features extraction; Secondly, the information Gain (IG) is used to select features to improve the interpretability of the model; Then, the proposed method was validated using data from the MPPT mode within the GPVS-faults and other dataset, and comparing the accuracy and calculation time of the MAFDT and other supervised learning models; Finally, the MAFDT model is deployed on the DSP for online fault diagnosis of grid-connected PV system. The results maintain 100% accuracy while cutting monitoring and diagnosis times to 0.0201 ms and 0.0364 ms, delivering 66-fold and 22-fold speed gains and high accuracy on the other dataset, which is especially suitable for online diagnosis of grid-connected PV system faults, and can ensure the reliability and safety of online operation of the system.

The cross-domain remaining useful life (RUL) prediction of lithium-ion batteries is critical for the safe and reliable operation of energy storage systems. However, practical applications still face two core challenges: limited real-time performance for online monitoring and reduced cross-domain generalization due to source-target distribution shifts. To address this issue, this paper proposes a transferable online state monitoring and RUL prediction method for lithium-ion batteries based on area ratio features. First, three area ratio features are extracted from the full charge-discharge curve, each exhibiting a correlation coefficient greater than 0.93 with the battery's capacity. Subsequently, a cycle compression and degradation state mapping strategy is applied to unify the time scale, and a cross-domain binary classification model is developed to enable online state monitoring. Furthermore, clustering analysis is conducted on rapidly degrading samples, and key subclasses are selected to construct the transfer model. Finally, a radial basis function (RBF) kernel-based maximum mean discrepancy (MMD) constraint is incorporated into the long short-term memory (LSTM) feature representation layer to align the feature distributions of the source and target domains. Experimental results demonstrate that the cross-domain classification accuracy of the online state monitoring model exceeds 93%, with a runtime of less than 0.5 ms. The transfer-based RUL prediction model achieves a mean absolute percentage error (MAPE) of under 4%, indicating that the proposed method combines high prediction accuracy with real-time performance.

Accurate prediction of the remaining useful life (RUL) of industrial equipment plays a crucial role in effective health management. Current deep learning methods fail to extract features across multiple time scales, limiting the model’s ability to comprehensively learn the equipment degradation patterns. To address these challenges, a novel encoder-decoder temporal convolutional network (TCN) based on the dual-region guided multi-head attention (DRGMA) mechanism is RUL prediction of rotating machinery. First, an interactive channel attention (ICA) module is designed in the encoder structure to highlight key degradation features while suppressing redundancy. Next, the DRGMA mechanism is introduced to guide the model in capturing degradation features across various scales, thereby enhancing both its performance and transparency. In the decoder structure, a TCN structure is constructed based on a prediction channel and a backtracking channel, which can capture forward dynamic patterns and residual information embedded within the input sequence. The former extracts critical temporal cues through cascaded causal convolutions, while the latter separates noise and redundancy by exploiting the differences between the input and the prediction. Experimental results on the PHM2012 and C-MAPSS datasets indicate that the proposed model outperforms existing methods in terms of RMSE and Score metrics.

Accurately and reliably predicting the remaining useful life (RUL) of lithium-ion batteries is crucial for determining the optimal time for battery replacement. However, many existing methods ignore the distribution shift problem between training and testing data across formation protocols and environmental temperature scenarios, leading to significant performance degradation in new scenarios. Therefore, this paper proposes a transfer prediction method based on DDARU (Dual-Domain Adversarial Residual Unscented Transfer). Firstly, a dualdomain adversarial neural network (DDANN) is proposed, which generates feature representations with crossdomain adaptability through the collaborative training of a multi-class domain discriminator and a dualdomain feature extractor. Secondly, a residual gated recurrent unit (Res-GRU) model is designed to enhance the time-dependent modeling ability of key degradation stages and achieve precise SOH prediction. Finally, the unscented particle filter algorithm (UPF) is adopted to dynamically correct the initial SOH prediction values, effectively eliminating noise interference and improving the tracking ability of nonlinear degradation processes. Experimental results on the University of Michigan Battery Lab dataset show that the proposed method improves the RMSE and R2 metrics by an average of about 5 % compared to the comparison methods, and demonstrates strong adaptability and stability across protocol and temperature scenarios.

Accurately predicting the remaining useful life (RUL) of lithium-ion batteries is essential for determining appropriate replacement times. Traditional RUL prediction methods often suffer from poor accuracy and robustness when batteries have undergone different formation protocols. To overcome these challenges, we propose a novel RUL prediction method that employs an Unscented Transformation tailored for variable formation protocols and a hybrid UPF-GRU framework with a protocol-sensitivity compensation module. First, we develop an exponential empirical degradation model using an Unscented Particle Filter (UPF) to capture the battery aging mechanism. Next, we construct a Gated Recurrent Unit (GRU) neural network that takes key features from constant-current/constant-voltage charge-discharge cycles as inputs and treats the battery’s state of health (SOH) as a hidden state, with RUL under varying formation protocols as the output. By integrating empirical modeling with data-driven prediction, our method refines the RUL estimates and adapts dynamically to protocol variations. We validate the approach on data from the University of Michigan Battery Lab. The results show that the proposed method achieves a coefficient of determination R² of 99.41

Due to the difference in the distribution of virtual and physical data caused by external influences such as load and rotational speed fluctuations, it is difficult for digital twins (DTs) to perform data augmentation and knowledge transfer, which results in low diagnostic accuracy for rolling bearings. To alleviate this problem, this study introduces an intelligent diagnostic framework that combines an adaptively updated DT with high-fidelity models and a dual-domain collaborative lightweight adversarial transfer network (DCLATN). This framework facilitates information transfer from physical entities to dynamic models, thereby enhancing cross-domain adaptability for fault diagnosis under complex operating conditions. Firstly, a high-fidelity DT model reflecting the physical operating state of the rolling bearing was established through dynamic modeling and structural parameter updating, and the model was then simulated and analyzed. Then, the DCLATN is constructed to minimize the discrepancy between the source and target domains, where the depth-separable convolution layers with weight-sharing mechanism are constructed for feature extraction and the gradient inversion layers are utilized for feature discrimination. The superiority of the proposed method lies in that it effectively eliminates the domain distribution disparities in vibration signals under diverse operating conditions, thereby ensuring the robustness and efficiency of the model in cross-domain fault diagnosis. The experimental results verify that the proposed method surpasses the state-of-the-art bearing fault diagnosis methods.

Insulated Gate Bipolar Transistor (IGBT) is recognized as having its Remaining Useful Life (RUL) prediction constitute a crucial component in the implementation of Prognostics and Health Management (PHM). The investigation of IGBT RUL under varying operating conditions is considered to possess significant theoretical importance and engineering value. A transfer learning-based RUL prediction method is proposed in this study. Initially, corresponding current and voltage signals are extracted from aging data according to the operational characteristics of IGBTs, from which the on-state resistance is calculated and characterized. Subsequently, features from both source and target domain data are systematically selected and fused to construct a health indicator with high transferability. Finally, a Maximum Mean Discrepancy-based Domain Adversarial Neural Network (MMD-DANN) is employed to minimize cross-domain discrepancies, thereby enabling high-accuracy RUL prediction to be achieved across different operating conditions. Experimental validation demonstrates that the proposed method is proven to provide an effective solution for reliability assessment of power devices under different operational conditions.

Feature extraction and fusion are important for the fault diagnosis and prediction of rotating machinery. While traditional deep learning networks can learn single attribute of features, they still face difficulties in capturing heterogeneous features with distinct attributes during the fusion process. To solve this problem, a novel type of capsule networks (CapsNets) based on dual heterogeneous feature resonance fusion is presented for heterogeneous feature extraction and fault diagnosis. Firstly, a dual-scale deformable convolution network is proposed to extract dual heterogeneous features. Then, an adaptive heterogeneous feature adjustment mechanism is presented to adjust the weights of heterogeneous features and identify discriminative features. Next, a resonance fusion mechanism is constructed to coordinate and select correlated heterogeneous features in both structural and spatial dimensions, avoiding information conflicts in feature fusion. Lastly, the heterogeneous resonance gain features are introduced into the CapsNet for fault diagnosis and classification tasks. The superiority of the proposed network lies in its ability to integrate and coordinate global and local information, enhancing the correlation between heterogeneous features for improved performance. Comparative experiments on multiple datasets with the state-of-the-art methods demonstrate that the proposed method excels in extracting and fusing dual heterogeneous features under complex operating conditions and noise interference.

The high-order spectral components of broadband signals interact with the PA memory effect, resulting in the inability of traditional models to effectively suppress intermodulation distortion (IMD). This paper proposes a novel cross-odd order augmented real-valued time-delay neural network Model (ARVTDNN), which has the potential to be used as an effective solution for digital predistortion (DPD) linearization in radio frequency (RF) power amplifiers (PA). This model clearly inhibits IMD caused by I/Q imbalance by introducing the cross-product term between the delayed I/Q component and the odd-order envelope parameter into the input vector. These vectors undergo certain delayed scutting, improving the basis functions and ultimately enhancing the modeling ability of the neural network. The experimental results show that the normalized mean square error (NMSE) of this model is 7 dB higher than that of long short-term memory (LSTM), 4 dB higher than that of ARVTDNN, 15 decibels higher than that of bidirectional long short-term memory (Bi-LSTM), 14.76 dB suppressed IMD3, and the coefficient is 45% lower than that of ARVTDNN. It shows excellent performance in reducing the NMSE and effectively suppressing intermodulation distortion.

Hyperspectral image (HSI) classification is a challenging task in remote sensing applications, aiming to determine the category of each pixel by utilizing rich spectral and spatial information in HSI. Convolutional neural networks (CNNs) have been effective in processing HSI data by extracting local features, but they are deficient in capturing global contextual information. Recently, transformer has become proficient in attending to global information due to their self-attention mechanisms, yet they may fall short in capturing multiscale features of HSI. To address these limitations, a global-local multigranularity transformer (GLMGT) network is proposed for HSI classification. The GLMGT combines CNN with the transformer to comprehensively capture multigranularity spectral and spatial features across global and local scales. Specifically, we introduce a multigranularity spatial feature extraction block to extensively extract spatial information at different granularities, including multiscale local spatial features and global spatial features. In addition, we introduce a multigranularity spectral feature extraction block to fully leverage spectral information across different granularities. The validity of the proposed method is demonstrated through experimental validation using seven publicly available datasets, which include two Chinese satellite hyperspectral datasets (ZY1-02D Huanghekou and GF-5 Yancheng) and one UAV-based hyperspectral dataset.

In the field of battery management, accurately predicting the remaining useful life (RUL) of lithium-ion batteries is critical to safeguard the safety of batteries and devices. However, current prediction methods are mainly based on sensor data such as battery voltage, current, temperature, time, and other data, and there is a challenge problem with neglecting the intrinsic connection between these data. To address this problem, a lithium-ion battery health state assessment and remaining useful life prediction method based on multi slope features of discharge curves is proposed. First, new slope features of discharge curve are proposed based on the discharge mechanism of lithium-ion batteries; second, slope features are utilized to estimate the state of health (SOH) of the batteries; then, unsupervised clustering is performed on the unlabeled SOH data to construct a health state assessment model; and finally, the RUL prediction model based on the gated recurrent unit with grey wolf optimizer (GWO-GRU) is constructed. The proposed method is validated using the aging data of lithium-ion batteries released by NASA Ames Research Center and the Oxford battery degradation dataset, and compared with other mainstream methods. The experimental results show that the method exhibits high accuracy in RUL prediction, the prediction accuracy is 99 % and 98 % respectively, which is significantly better than traditional methods. Therefore, the method can effectively assess the health state of lithium-ion batteries and accurately predict their RUL, which can provide support for the realization of lithium-ion battery health management and intelligent operation and maintenance.

Due to harsh working conditions and variable load, the vibration signal of the rolling bearing possess nonstationary and nonlinear properties, leading to low robustness and unreliability in fault diagnosis under noise interference. To this end, a uniform phase decoupled iterative filtering (IF) and multi-scale sliding fractal box dimension (UPDIF-MSFBD) is proposed, where UPDIF is employed to reduce the noise interference of vibration signal, and MSFBD is introduced to capture the multiscale fractal temporal features. First, a narrowband sinusoidal wave with a uniform phase distribution in IF is added into the original signal to equalize the distribution of extreme points of the intrinsic mode functions (IMF), where a new kind of decoupled IMFs is obtained based on the orthogonality between the IMFs to overcome the mode mixing problem. Second, the interquartile range normalization and the sliding coarse-grained method are introduced to explore abundant coarse-grained information and capture the fractal temporal features of fault vibration signals. Third, the extracted features are input into the classifiers for bearing fault diagnosis. The advantage of the UPDIF-MSFBD method is the robustness and effectiveness in the extraction and separation of multiscale fractal temporal features. The findings reveal that the proposed approach efficiently capture fault characteristics, outperforming current methods in terms of accuracy in fault diagnosis. The method is especially valuable for real-time monitoring and predictive maintenance, offering an advanced solution for fault diagnosis in complex machinery systems, which is crucial for minimizing downtime and extending the lifespan of critical components.

Compared to controlled laboratory conditions, the electrical performance of photovoltaic (PV) modules under real operating conditions is influenced by the nonlinear effects of dynamic environmental factors. Improving voltage parameter accuracy, especially Maximum Power Point Voltage (Vmp), is crucial for efficient Maximum Power Point Tracking (MPPT) and overall system performance. Traditional models that rely solely on module temperature and irradiance fail to adequately capture outdoor climate variations. This paper proposes an analytical Model for predicting voltage parameters based on the temperature difference between the module and ambient(Delta T), considering their interaction. First, the strong nonlinear correlation between Delta T and voltage parameters was determined using correlation analysis methods such as the Maximal Information Coefficient (MIC). Then, Delta T was quantified and integrated into the traditional conversion formula. Finally, the formula coefficients were identified using the levenberg-marquardt (L-M) method with limited historical data. The proposed model was validated using data from nine PV module groups with six technologies under various climates from two public datasets. Results show that incorporating the Delta T formula improves prediction accuracy and environmental adaptability of traditional model. The RMSE for Open Circuit Voltage (Voc) decreased by 0.1050 to 0.6389 V and for Vmp by 0.1004 to 1.2484 V, with the reduction in error being more especially significant under high-temperature conditions above 40 degrees C. The identified coefficients show good stability and consistency. Furthermore, predictions for the same PV modules deployed under different climate conditions validated the model's good generalization ability.

Feature extraction remains a challenging task in bearing fault diagnosis due to the presence of nonlinearity, nonstationarity, and noise interference. To address this issue, an extended iterative filtering and composite multiscale fractional-order Boltzmann-Shannon interaction entropy (EIF-CMFBSIE) are proposed for for rolling bearing fault diagnosis in complex environments. First, an EIF method is proposed to decompose the vibration signal into multiple intrinsic mode functions (IMFs) by extending the lengths of both ends of the signal through waveform matching. Second, multi-scale coarse-graining is applied to each IMF, fractional-order BoltzmannShannon interaction entropy (FBISE) is computed for each coarse-grained sequence by incorporating fractionalorder parameters, and CMFBSIE is obtained through composite averaging to construct a multi-dimensional fault feature set. Next, the joint approximate diagonalization of eigenmatrices (JADE) method is employed to eliminate redundant information and fuse the fault features. The fused feature sets are then input into the kernel extreme learning machine (KELM) classifier for multi-fault identification. The proposed EIF-CMFBSIE method demonstrates excellent performance in analyzing the nonlinear dynamic complexity and irregularity of vibration signals in noisy environments. In the fault diagnosis tests based on three bearing simulation test benches, compared with the existing five fault diagnosis methods, the recognition accuracy of EIF-CMFBSIE is increased by 13.33%, and there is a significant advantage in computational efficiency, in which the EIF shortens the decomposition time by 65-96% compared with the existing methods. The experimental results indicate that the method can not only accurately identify different fault types and the degree of faults, but also has a short calculation time and better overall performance.

State of health estimation of lithium-ion batteries is essential for ensuring operational safety, extending service life, and optimizing production processes. To address the limitation of relying solely on features from either the charging or discharging process, this study proposes a health state estimation method based on features extracted from both charging and discharging phases. Specifically, three features are extracted from the battery's charging and discharging curves and then evaluated for their correlation with SOH using the spearman correlation coefficient to ensure the effectiveness. A particle swarm optimization-optimized gate recurrent unit-based model is constructed to estimate state of health, and the proposed method is validated using the CALCE dataset and NASA dataset. Experimental results show that the root mean square error is controlled to around 1%, demonstrating that the method can effectively and accurately assess the SOH of lithium-ion batteries, providing valuable guidance for practical applications in battery management systems.

Condition monitoring and state of health estimation are crucial for ensuring the secure operation of lithium-ion batteries. To address the difficulty in thoroughly capturing the changes in the state of the partial charging and discharging processes of the battery and the high demand for real-time performance of battery condition monitoring, a method for online condition monitoring and state of health estimation of lithium-ion batteries based on time-ratio features from the whole process is proposed. The health features are derived from the battery's full charging and discharging curves, and the correlation methodology is used to analyse the relationship with the state of health. The proposed random forest-based online condition monitoring model and gated recurrent unit-based state of health estimation model are conducive to timely monitoring abnormal conditions and accurate estimation to ensure safe battery operation. Finally, the proposed method can realise accurate and real-time online condition monitoring with a classification accuracy of more than 0.93 and running time of less than 0.2 ms, and achieve better performance of state of health estimation under different ambient temperatures, with a root-mean-square error of less than 0.02 at room temperature and less than 0.01 at 43 degrees C.

Accurate prediction of photovoltaic (PV) power generation is crucial for enhancing the efficiency of power system dispatch and the integration capacity of renewable energy, holding significant engineering implications for grid stability and management. This study proposes a photovoltaic power prediction method based on meteorological data and backplane temperature. First, a Pearson correlation analysis is conducted to identify and select input features with high relevance to power output. Second, a genetic algorithm-optimized neural network model is developed to predict backplane temperature, which captures the complex nonlinear relationships between environmental variables and component thermal behavior. Third, the predicted backplane temperature is integrated as a key physical input, along with other selected variables such as irradiance and ambient temperature, into a neural network model that employs time-series analysis and feature fusion for final power forecasting. Experimental results demonstrate that this structured three-step approach significantly enhances the accuracy and robustness of the power prediction model.

Entropy-based feature extraction methods have emerged as a focal point in fault diagnosis, leveraging their inherent advantages, which include independence from prior knowledge, elimination of the need for preprocessing, and straightforward implementation. However, multiscale entropy encounters challenges in capturing early fault characteristics, as it predominantly emphasizes low-frequency fault information, potentially overlooking valuable high-frequency data. To address this limitation, a novel hierarchical fractional-order Boltzmann–Shannon interaction entropy is proposed to extract fault information across both high and low frequencies and enhance the noise-resistant performance of the algorithm. By hierarchically decomposing vibration signals and subsequently computing the fractional-order Boltzmann–Shannon interaction entropy of the resultant sub-signals at each level, a comprehensive multidimensional fault characteristic vector is constructed. Furthermore, to mitigate the sensitivity to parameter selection within the kernel extreme learning machine (KELM) model, this study incorporates the Newton–Raphson-based optimizer (NRBO) to optimize the regularization coefficients and kernel function parameters, thereby establishing an optimal NRBO–KELM model for gear fault diagnosis. Validation with the WT-planetary gearbox datasets and HUST gearbox fault datasets demonstrates that the proposed method achieves 100

The global modeling capability of vision transformer (ViT) has been well proven in the field of hyperspectral image (HSI) classification. However, ViT does not have the excellent local feature extraction capability compared with the convolutional neural network (CNN). Therefore, early-stage convolutions are often used to enhance ViT's local representation ability. However, directly applying convolutions on high-dimensional HSI data increases computational overhead. Moreover, recent researches have observed that ViT may suffer from high redundancy in capturing multihead self-attention (MHSA). To address the above issues, we propose a multiscale super token transformer (MSSTT) model for HSI classification. We use a divide-and-conquer strategy to extract local features and global dependencies of HSI data at multiple granularities. Specifically, our proposed model incorporates two branches: a multiscale convolution (MSConv) branch that uses various convolutional kernels to extract diverse local features and a multiscale super token attention (MSSTA) branch for capturing global features with low redundancy. Finally, comparative experimental results with advanced methods show that the proposed MSSTT possesses better classification performance. On the Salinas (SA), Pavia University (PU), and Kennedy Space Center (KSC) datasets, the overall accuracies (OAs) of our MSSTT are 98.47%, 98.47%, and 99.38%, respectively. Code will be released at https://github.com/zhangtaizheng/MSSTT.

IGBT (insulated gate bipolar transistor) undergo aging failures due to environmental effects during operation. It is of theoretical importance and engineering value to study how to predict the remaining useful life (RUL) of IGBT using multi-source data. In this paper, we propose a multi-source data-driven IGBT RUL prediction method, which extracts the time-domain features of the degradation process by utilizing the operating characteristics of IGBT and employs a bi-directional gated recurrent unit neural network to improve the prediction performance. First, the corresponding current-voltage signals are extracted from the aging data according to the IGBT operating characteristics, and the time-domain features are calculated; second, the random forest algorithm is used to rank the importance of the features, and the important features are screened out to reduce the redundancy of the features; finally, the RUL prediction model based on the BiGRU network is established, and the validity of the method is verified by the actual IGBT degradation feature data. The results show that the proposed method combines a variety of aging data, fully considers multiple factors in the working condition, and has higher RUL prediction accuracy and stability.